Automatic sprinkler system and method based on multi-modal perception and intelligent decision making
The automatic sprinkler fire extinguishing system, which integrates multimodal perception and intelligent decision-making, utilizes a hybrid deep learning model to fuse temperature and smoke data, enabling rapid and accurate fire identification. This solves the problems of response delay and false alarms in traditional systems, achieving efficient fire identification and suppression.
Patent Information
- Application Number
- CN202610273308.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-06
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional automatic sprinkler fire extinguishing systems have delayed response and are susceptible to environmental interference, making it difficult to quickly and accurately identify initial fires. Existing multi-sensor systems lack deep fusion of multi-modal signals and intelligent decision-making capabilities.
An automatic sprinkler fire extinguishing system employing multimodal perception and intelligent decision-making combines temperature sensing and smoke visual sensing, utilizes a hybrid deep learning model for data processing, including convolutional neural networks and long short-term memory networks, to perform feature fusion and decision-making, and generate fire extinguishing trigger commands.
It achieves a fire identification response speed of less than 3 seconds, an accuracy rate of over 99%, and a false alarm rate of less than 0.7%. It effectively distinguishes between real fires and interference in complex environments, thereby improving the success rate of initial fire suppression.
Smart Images
Figure CN122424540A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of fire safety technology, and in particular to an automatic sprinkler fire extinguishing system and method based on multimodal perception and intelligent decision-making. Background Technology
[0002] With the widespread adoption of new energy vehicles, emerging fires such as lithium battery fires are characterized by rapid ignition, rapid temperature rise, and difficulty in extinguishing. Traditional automatic sprinkler systems mostly rely on single temperature sensors, resulting in response delays often exceeding 30 seconds. They are also susceptible to interference from environmental factors such as kitchen fumes and vehicle exhaust, leading to high false alarm rates and failing to meet the need for rapid suppression of initial fires. While some existing technologies attempt to combine multiple sensors, they primarily employ simple threshold judgments, lacking the ability to deeply fuse and intelligently understand multimodal signals. This hinders the achievement of high-accuracy, low-latency fire identification in complex environments. Therefore, there is an urgent need for an automatic fire extinguishing system that can integrate multi-sensor information, possess intelligent learning and decision-making capabilities, and offer rapid and reliable response.
[0003] Application content
[0004] This invention aims to solve the above-mentioned problems by proposing an automatic sprinkler fire extinguishing system and method based on multimodal perception and intelligent decision-making, which features fast response speed, high recognition accuracy, strong environmental adaptability, and forward-looking early warning capabilities.
[0005] The technical solution of this invention is implemented as follows:
[0006] An automatic sprinkler fire suppression system based on multimodal perception and intelligent decision-making, including
[0007] Sensing module: Used to collect real-time environmental data of the monitoring area, including temperature time series data and smoke visual data; temperature time series data is collected through temperature sensing unit, and smoke images or smoke concentration distribution smoke visual data are collected through smoke visual sensing unit.
[0008] Data processing and control module: Communicatively connected to the sensing module, the data processing and control module includes a signal preprocessing unit and a fire intelligent identification unit. The signal preprocessing unit is used to standardize the data collected by the sensing module; the fire intelligent identification unit has a trained hybrid deep learning model embedded in it, which is used to output the probability of fire occurrence based on the data collected by the sensing module.
[0009] The hybrid deep learning model includes a convolutional neural network branch for extracting visual features of smoke, a long short-term memory network branch for extracting temporal features of temperature, and a feature fusion module for feature fusion and decision-making.
[0010] The decision-making and triggering unit is used to generate a fire extinguishing triggering command based on the fire occurrence probability output by the fire intelligent identification unit and a preset incremental logic judgment strategy.
[0011] Execution module: Connected to the data processing and control module, used to execute sprinkler fire extinguishing operation in response to the fire extinguishing trigger command.
[0012] Preferably, the execution module includes:
[0013] The spray unit includes high-pressure fine water mist nozzles;
[0014] The water supply unit is used to provide a constant pressure water source for the sprinkler unit;
[0015] The alarm communication unit is used to send alarm information synchronously when fire extinguishing is triggered.
[0016] Preferably, it also includes a system health management module, which is used to monitor the status of each module of the system, perform fault self-diagnosis and protection.
[0017] Preferably, the convolutional neural network branch includes at least two convolutional layers and pooling layers for extracting spatial features from the smoke image; the long short-term memory network branch includes LSTM layers and attention mechanism layers for extracting time-dependent feature patterns from temperature time-series data; the feature fusion module concatenates the feature vectors from the two branches and maps them to a fire probability value through a fully connected network.
[0018] Preferably, the incremental logic judgment strategy executed by the decision and triggering unit is as follows: a fire extinguishing trigger command is generated only when the fire occurrence probability value exceeds the first threshold for N consecutive sampling periods, and exceeds the second threshold higher than the first threshold at least once during the time period; where N is an integer greater than 1.
[0019] Preferably, the training method for the hybrid deep learning model includes:
[0020] Constructing a training dataset: The dataset contains temperature sequences and smoke image samples collected synchronously under various real fire scenarios (such as lithium battery thermal runaway and solid material combustion) and typical interference scenarios (such as cooking smoke, vehicle exhaust, and steam), and each sample is labeled as "fire" or "non-fire".
[0021] Model training: Using the training dataset, with the temperature sequence and smoke image as input, and the label as supervision signal, the Adam optimization algorithm is used to train the hybrid deep learning model end-to-end with the goal of minimizing the binary cross-entropy loss. Dropout technique is used to prevent overfitting during training, and early stopping is used to determine the optimal number of training rounds based on the performance of the validation set.
[0022] This invention also provides an automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making, applied to the above-mentioned system, comprising the following steps:
[0023] S1: Continuously collects time-series temperature data and visual smoke data of the monitored area through the sensing module;
[0024] S2: Preprocess the collected data;
[0025] S3: Input the preprocessed multimodal data into a preset hybrid deep learning model. The hybrid deep learning model processes the smoke visual data through its convolutional neural network branch to extract visual features, processes the temperature time series data through its long short-term memory network branch to extract time series features, and outputs the probability of fire occurrence after fusing the two types of features.
[0026] S4: Based on the fire occurrence probability and combined with the incremental logic judgment strategy, determine whether the fire extinguishing trigger condition is met;
[0027] S5: If the triggering condition is met, the control execution module will start the sprinkler fire suppression system and send an alarm message.
[0028] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making as described above.
[0029] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making.
[0030] By adopting the above technical solution, the beneficial effects of the present invention are as follows:
[0031] 1. By combining multimodal perception with efficient hybrid deep learning model inference, the decision-making time from fire identification to initial response is shortened to less than 3 seconds, far exceeding the 30 seconds or more of traditional systems, which greatly improves the success rate of extinguishing initial fires.
[0032] 2. By leveraging the deep understanding of smoke morphology by convolutional neural networks and the ability of LSTM to capture temperature change trends, combined with the attention mechanism to focus on key information, it can effectively distinguish between real fires and complex environmental interference. Experimental verification shows that the false alarm rate is less than 0.7% and the accuracy rate exceeds 99%.
[0033] 3. The multimodal data provided by the sensing module and the intelligent recognition algorithm form a functional closed loop. The algorithm relies on sensor data to make accurate judgments, and the algorithm's fast and accurate decision-making improves the execution efficiency of the entire system. The two support and interact with each other, and together solve the technical problem of rapid and accurate fire extinguishing. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0035] Figure 1 This is a structural block diagram of the present invention;
[0036] Figure 2 This is a schematic diagram of the architecture of the hybrid deep learning model in this invention;
[0037] Figure 3 This is a flowchart illustrating the intelligent fire identification and decision-making process in this invention;
[0038] Figure 4 This is a histogram comparing the performance of the traditional solution and the solution of this invention on the test set. Detailed Implementation
[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0040] like Figure 1 As shown, the automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making includes a perception module, a data processing and control module, and an execution module.
[0041] The sensing module collects temperature time-series data through the temperature sensing unit and collects smoke images or smoke concentration distribution visual data through the smoke visual sensing unit. The temperature sensor uses a DS18B20 digital sensor, and the smoke visual sensor uses an OV2640 camera with an infrared filter to enhance the sensitivity to combustion particles.
[0042] The data processing and control module is communicatively connected to the sensing module. The data processing and control module is an embedded AI computing chip (NVIDIA Jetson Nano is used in this application), on which a fire intelligent identification software unit is deployed. The fire intelligent identification software unit includes a signal preprocessing unit and a fire intelligent identification unit. The signal preprocessing unit is used to standardize the data collected by the sensing module. The fire intelligent identification unit has a trained hybrid deep learning model embedded within it, used to output the probability of fire occurrence based on the data collected by the sensing module.
[0043] like Figure 2 As shown, the hybrid deep learning model includes a convolutional neural network branch for extracting visual features of smoke, a long short-term memory network branch for extracting time-series temperature features, and a feature fusion module for feature fusion and decision-making.
[0044] Model structure:
[0045] Convolutional Neural Network Branch: Input is a 64x64 grayscale smoke image. First layer: 32 3x3 convolutional kernels, ReLU activation, 2x2 max pooling; Second layer: 64 3x3 convolutional kernels, ReLU activation, 2x2 max pooling; then flattened into a 1024-dimensional feature vector.
[0046] Long Short-Term Memory (LSTM) branch: The input is a standardized temperature sequence of length 60. The layer structure is a 128-unit bidirectional LSTM layer followed by an attention layer, and then a fully connected layer with 64 neurons (ReLU activation). The output is a 64-dimensional feature vector.
[0047] Feature fusion module: The above 1024-dimensional and 64-dimensional vectors are concatenated and input into a fully connected layer with 256 neurons (ReLU activation, Dropout=0.5), then through a fully connected layer with 64 neurons (ReLU activation), and finally through a sigmoid-activated output neuron to obtain the fire probability P.
[0048] Model training methods:
[0049] Over 100,000 labeled sample pairs (temperature sequences, smoke images) were collected, with a fire scene to interference scene ratio of approximately 1:4. 20-30% of the data samples underwent augmentation, including random image rotation (±15°) and brightness adjustment. The PyTorch framework was used with a batch size of 32, an initial learning rate of 0.001, the Adam optimizer, and the ReduceLROnPlateau scheduler. Training was stopped early after approximately 100 epochs. The trained model was then quantized and optimized before deployment to an embedded chip. Inference latency was less than 300 milliseconds.
[0050] like Figure 3 As shown, the decision and triggering unit is used to generate a fire extinguishing triggering command based on the fire occurrence probability output by the fire intelligent identification unit and a preset incremental logic judgment strategy. The incremental logic judgment strategy executed by the decision and triggering unit is as follows: when the fire occurrence probability value P exceeds the first threshold (TH1=0.7) for three consecutive sampling periods (once per second), and exceeds the second threshold (TH2=0.85) higher than the first threshold at least once during the period, it is determined to be a fire and fire extinguishing is triggered.
[0051] Execution module: connected to the data processing and control module, used to execute the spray fire extinguishing operation in response to the fire extinguishing trigger command. The execution module includes: a spray unit containing high-pressure fine water mist nozzles, a water supply unit for providing a constant pressure water source to the spray unit, and an alarm communication unit for synchronously sending alarm information when fire extinguishing is triggered.
[0052] As another preferred embodiment of the present invention, it also includes a system health management module, which is used to monitor the status of each module of the system, perform fault self-diagnosis and protection.
[0053] In another preferred embodiment of the present invention, the convolutional neural network branch includes at least two convolutional layers and pooling layers for extracting spatial features from smoke images; the long short-term memory network branch includes LSTM layers and attention mechanism layers for extracting time-dependent feature patterns from temperature time-series data; the feature fusion module concatenates the feature vectors from the two branches and maps them to fire probability values through a fully connected network.
[0054] Application Scenarios and Effect Verification
[0055] This system was installed in the underground elevated area where electric bicycles are centrally parked in the residential community. During six months of operation, the system recorded and processed over 2.6 million detection events. It successfully responded to and extinguished two initial open flames caused by battery malfunctions within 3 seconds, effectively preventing the fires from spreading. Simultaneously, the system did not generate false alarms due to daily disturbances such as vehicle dust and smoking, with a measured false alarm rate of 0.05%. Compared to the traditional temperature-sensitive sprinkler system previously deployed in the same area (which had been falsely triggered by steam), this system demonstrated significant advantages in accuracy and reliability.
[0056] Comparative Example: In a controlled laboratory environment, a test set containing 2000 samples was constructed to compare the performance of the system of this invention (using the aforementioned intelligent recognition algorithm) with two traditional methods (single temperature threshold method, temperature-smoke dual threshold method, and logic method). Results are as follows... Figure 4 And as shown in the table below:
[0057] Traditional single temperature threshold method 82.5% 15.8% 1.7% <1 second Traditional dual threshold and logic method 88.2% 10.1% 1.7% <1 second Invention Solution 99.1% 0.7% 0.2% <0.3 seconds
[0058] Data shows that this invention achieves a significant leap in technical performance by deeply integrating technical features (multimodal sensors) and algorithmic features (hybrid deep learning models), solving the technical challenge of rapid and accurate identification.
[0059] Components not described in detail in this article are existing technologies.
[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making, characterized in that: include Sensing module: Used to collect real-time environmental data of the monitoring area, including time-series temperature data and visual smoke data; Data processing and control module: Communicatively connected to the sensing module, the data processing and control module includes a signal preprocessing unit and a fire intelligent identification unit. The signal preprocessing unit is used to standardize the data collected by the sensing module; the fire intelligent identification unit has a trained hybrid deep learning model embedded in it, which is used to output the probability of fire occurrence based on the data collected by the sensing module. The hybrid deep learning model includes a convolutional neural network branch for extracting visual features of smoke, a long short-term memory network branch for extracting temporal features of temperature, and a feature fusion module for feature fusion and decision-making. The decision-making and triggering unit is used to generate a fire extinguishing trigger command based on the fire occurrence probability and a preset incremental logic judgment strategy. Execution module: Used to execute sprinkler fire extinguishing operation in response to the fire extinguishing trigger command.
2. The automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making according to claim 1, characterized in that: The execution module includes: The spray unit includes high-pressure fine water mist nozzles; The water supply unit is used to provide a constant pressure water source for the sprinkler unit; The alarm communication unit is used to send alarm information synchronously when fire extinguishing is triggered.
3. The automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making according to claim 1, characterized in that: It also includes a system health management module, which is used to monitor the status of each module of the system, perform fault self-diagnosis and protection.
4. The automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making according to claim 1, characterized in that: The convolutional neural network branch contains at least two convolutional layers and pooling layers for extracting spatial features from smoke images; The Long Short-Term Memory (LSTM) network branch includes an LSTM layer and an attention mechanism layer, used to extract time-dependent feature patterns from temperature time-series data; the feature fusion module concatenates the feature vectors from the two branches and maps them to fire probability values through a fully connected network.
5. The automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making according to claim 1, characterized in that: The incremental logic judgment strategy executed by the decision and triggering unit is as follows: a fire extinguishing trigger command is generated only when the fire occurrence probability value exceeds the first threshold for N consecutive sampling periods, and exceeds the second threshold higher than the first threshold at least once during the time period; where N is an integer greater than 1.
6. The automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making according to claim 1, characterized in that: The training methods for the hybrid deep learning model include: Constructing a training dataset: The dataset contains temperature sequences and smoke image samples collected synchronously under various real fire scenarios and typical interference scenarios, and each sample is labeled with a fire or non-fire label; Model training: Using the training dataset, with the temperature sequence and smoke image as input, and the label as supervision signal, the Adam optimization algorithm is used to train the hybrid deep learning model end-to-end with the goal of minimizing the binary cross-entropy loss. Dropout technique is used to prevent overfitting during training, and early stopping is used to determine the optimal number of training rounds based on the performance of the validation set.
7. An automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making, applied to the automatic sprinkler fire extinguishing system based on multimodal perception and intelligent decision-making as described in any one of claims 1-6, characterized in that: Includes the following steps: S1: Continuously collects time-series temperature data and visual smoke data of the monitored area through the sensing module; S2: Preprocess the collected data; S3: Input the preprocessed multimodal data into a preset hybrid deep learning model. The hybrid deep learning model processes the smoke visual data through its convolutional neural network branch to extract visual features, processes the temperature time series data through its long short-term memory network branch to extract time series features, and outputs the probability of fire occurrence after fusing the two types of features. S4: Based on the fire occurrence probability and combined with the incremental logic judgment strategy, determine whether the fire extinguishing trigger condition is met; S5: If the triggering condition is met, the control execution module will start the sprinkler fire suppression system and send an alarm message.
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, it implements the steps of the automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making as described in claim 7.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by the processor, it implements the steps of the automatic sprinkler fire extinguishing method based on multimodal perception and intelligent decision-making as described in claim 7.